Dataset


The stereo / flow benchmark consists of 194 training image pairs and 195 test image pairs, saved in loss less png format. Our evaluation server computes the average number of bad pixels for all non-occluded or occluded (=all groundtruth) pixels. We require that all methods use the same parameter set for all test pairs. Our development kit provides details about the data format as well as MATLAB / C++ utility functions for reading and writing disparity maps and flow fields.

Evaluation

Our evaluation table ranks all methods according to the number of non-occluded erroneous pixels at the specified disparity / end-point error threshold. All methods providing less than 100 % density have been interpolated using simple background interpolation as explained in the corresponding header file in the development kit. For each method we show:

  • Out-Noc: Percentage of erroneous pixels in non-occluded areas
  • Out-All: Percentage of erroneous pixels in total
  • Avg-Noc: Average disparity / end-point error in non-occluded areas
  • Avg-All: Average disparity / end-point error in total
  • Density: Percentage of pixels for which ground truth has been provided by the method

Note: Our main ranking is computed at 3 pixels error threshold, evaluating all pixels. For methods which do not provide dense result we use background interpolation to fill in missing values.

  • Stereo: Method uses left and right (stereo) images
  • Flow: Method uses optical flow (2 temporally adjacent images)
  • Multiview: Method uses more than 2 temporally adjacent images
  • Laser Points: Method uses point clouds from Velodyne laser scanner
  • GPS: Method uses GPS information
  • Motion stereo: Method uses epipolar geometry for computing optical flow

Error threshold        Evaluation area

Optical Flow Evaluation

This table ranks general optical flow methods, performing a full 2D search, as compared to the motion stereo methods below.

Rank Method Setting Code Out-Noc Out-All Avg-Noc Avg-All Density Runtime Environment
1 SceneFlow
This method uses stereo information.
This method uses optical flow information.
This method makes use of the epipolar geometry.
3.42 % 6.41 % 0.8 px 1.3 px 100.00 % 6 min 4 cores @ 3.0 Ghz (Matlab + C/C++)
Anonymous submission
2 PR-Sf+E
This method uses stereo information.
This method uses optical flow information.
4.08 % 7.79 % 0.9 px 1.7 px 100.00 % 200 s 4 cores @ 3.0 Ghz (Matlab + C/C++)
C. Vogel, S. Roth and K. Schindler: Piecewise Rigid Scene Flow. International Conference on Computer Vision (ICCV) 2013.
3 PCBP-Flow
This method uses optical flow information.
This method makes use of the epipolar geometry.
4.08 % 8.70 % 0.9 px 2.2 px 100.00 % 3 min 4 cores @ 2.5 Ghz (Matlab + C/C++)
K. Yamaguchi, D. McAllester and R. Urtasun: Robust Monocular Epipolar Flow Estimation. CVPR 2013.
4 PR-Sceneflow
This method uses stereo information.
This method uses optical flow information.
4.24 % 8.08 % 1.2 px 2.8 px 100.00 % 150 sec 4 core @ 3.0 Ghz (Matlab + C/C++)
C. Vogel, S. Roth and K. Schindler: Piecewise Rigid Scene Flow. International Conference on Computer Vision (ICCV) 2013.
5 MotionSLIC
This method uses optical flow information.
This method makes use of the epipolar geometry.
4.36 % 10.91 % 1.0 px 2.7 px 100.00 % 11 s 1 core @ 3.0 Ghz (C/C++)
K. Yamaguchi, D. McAllester and R. Urtasun: Robust Monocular Epipolar Flow Estimation. CVPR 2013.
6 gtRF-DF
This method uses optical flow information.
6.50 % 13.43 % 1.6 px 4.3 px 100.00 % 1 min 1 core @ 2.5 Ghz (Matlab + C/C++)
Anonymous submission
7 TGV2ADCSIFT
This method uses optical flow information.
6.55 % 15.35 % 1.6 px 4.5 px 100.00 % 12s GPU @ 2.4 Ghz (C/C++)
8 Data-Flow
This method uses optical flow information.
7.47 % 14.85 % 1.9 px 5.5 px 100.00 % 3 min 2 cores @ 2.5 Ghz (Matlab + C/C++)
C. Vogel, S. Roth and K. Schindler: An Evaluation of Data Costs for Optical Flow. German Conference on Pattern Recognition (GCPR) 2013.
9 DeepFlow
This method uses optical flow information.
7.49 % 17.93 % 1.5 px 5.8 px 100.00 % 17 s 1 core @ 3.6Ghz (Python + C/C++)
P. Weinzaepfel, J. Revaud, Z. Harchaoui and C. Schmid: DeepFlow: Large displacement optical flow with deep matching. IEEE Intenational Conference on Computer Vision (ICCV) 2013.
10 TVL1-HOG
This method uses optical flow information.
8.31 % 19.21 % 2.0 px 6.1 px 100.00 % 180 s 2 cores @ 3.0 Ghz (Matlab)
H. Rashwan, M. Mohamed, M. Garcia, B. Mertsching and D. Puig: Illumination Robust Optical Flow Model Based on Histogram of Oriented Gradients. German Conference on Pattern Recognition 2013 .
11 MLDP-OF
This method uses optical flow information.
8.91 % 18.95 % 2.5 px 6.7 px 100.00 % 160 s 2 cores @ 2.5 Ghz (Matlab)
Anonymous submission
12 DescFlow
This method uses optical flow information.
9.07 % 19.64 % 2.1 px 5.7 px 100.00 % 9.0 s GPU @ 2.5 Ghz (C/C++)
Anonymous submission
13 CRTflow
This method uses optical flow information.
9.71 % 18.88 % 2.7 px 6.5 px 100.00 % 18 s GPU @ 1.0 Ghz (C/C++)
O. Demetz, D. Hafner and J. Weickert: The Complete Rank Transform: A Tool for Accurate and Morphologically Invariant Matching of Structure. Proc.~British Machine Vision Conference 2013 (BMVC) 2013.
14 C++
This method uses optical flow information.
code 10.16 % 20.29 % 2.6 px 7.1 px 100.00 % 8.5 min 1 core @ 3.0 Ghz (Matlab)
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and The Principles Behind Them. 2013.
15 C+NL
This method uses optical flow information.
code 10.60 % 20.66 % 2.8 px 7.2 px 100.00 % 14.8 min 1 core @ 3.0 Ghz (Matlab)
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and The Principles Behind Them. 2013.
16 IVANN
This method uses optical flow information.
10.85 % 21.17 % 2.7 px 7.4 px 100.00 % 1073 s 1 core @ 2.5 Ghz (Matlab)
Anonymous submission
17 fSGM
This method uses optical flow information.
11.03 % 22.90 % 3.2 px 12.2 px 100.00 % 60 s 1 core @ 2.4 Ghz (C/C++)
S. Hermann and R. Klette: Hierarchical Scan Line Dynamic Programming for Optical Flow using Semi-Global Matching. ACCV Workshops 2012.
18 TGV2CENSUS
This method uses optical flow information.
code 11.14 % 18.42 % 2.9 px 6.6 px 100.00 % 4 s GPU+CPU @ 3.0 Ghz (Matlab + C/C++)
M. Werlberger: Convex Approaches for High Performance Video Processing. 2012.
R. Ranftl, S. Gehrig, T. Pock and H. Bischof: Pushing the Limits of Stereo Using Variational Stereo Estimation. IV 2012.
19 C+NL-fast
This method uses optical flow information.
code 12.42 % 22.27 % 3.2 px 7.8 px 100.00 % 2.9 min 1 core @ 3.0 Ghz (Matlab)
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and The Principles Behind Them. 2013.
20 EPPM
This method uses optical flow information.
code 13.04 % 23.76 % 2.5 px 9.2 px 100.00 % 0.25 s GPU @ 1.0 Ghz (C/C++)
Anonymous submission
21 HS
This method uses optical flow information.
code 14.77 % 24.08 % 4.0 px 9.0 px 100.00 % 2.6 min 1 core @ 3.0 Ghz (Matlab)
D. Sun, S. Roth and M. Black: A Quantitative Analysis of Current Practices in Optical Flow Estimation and The Principles Behind Them. 2013.
22 IQFlow
This method uses optical flow information.
18.93 % 28.33 % 3.6 px 8.8 px 100.00 % 60 s 4 cores @ 3.5 Ghz (C/C++)
Anonymous submission
23 GC-BM-Bino
This method uses stereo information.
This method uses optical flow information.
This method makes use of the epipolar geometry.
18.93 % 29.37 % 5.0 px 12.0 px 83.73 % 1.3 s 2 cores @ 2.5 Ghz (C/C++)
B. Kitt and H. Lategahn: Trinocular Optical Flow Estimation for Intelligent Vehicle Applications. ITSC 2012.
24 C+NL-M
This method uses optical flow information.
19.17 % 26.35 % 7.4 px 14.5 px 100.00 % 5 min 2 cores @ 2.5 Ghz (Matlab)
Anonymous submission
25 eFolki
This method uses optical flow information.
19.34 % 28.79 % 5.2 px 10.8 px 100.00 % 0.026 s GPU @ 700 Mhz (C/C++)
Anonymous submission
26 GC-BM-Mono
This method uses optical flow information.
This method makes use of the epipolar geometry.
19.49 % 29.88 % 5.0 px 12.1 px 84.33 % 1.3 s 2 cores @ 2.5 Ghz (C/C++)
B. Kitt and H. Lategahn: Trinocular Optical Flow Estimation for Intelligent Vehicle Applications. ITSC 2012.
27 RSRS-Flow
This method uses optical flow information.
20.74 % 29.68 % 6.2 px 12.1 px 100.00 % 4 min 1 core @ 2.5 Ghz (Matlab)
P. Ghosh and B. Manjunath: Robust Simultaneous Registration and Segmentation with Sparse Error Reconstruction. PAMI 2012.
28 ALD
This method uses optical flow information.
21.35 % 30.65 % 10.9 px 16.0 px 100.00 % 110 s 1 core @ 2.5 Ghz (C/C++)
M. Stoll, S. Volz and A. Bruhn: Adaptive Integration of Feature Matches into Variational Optical Flow Methods. ACCV 2012.
29 LDOF
This method uses optical flow information.
code 21.86 % 31.31 % 5.5 px 12.4 px 100.00 % 1 min 1 core @ 2.5 Ghz (C/C++)
T. Brox and J. Malik: Large Displacement Optical Flow: Descriptor Matching in Variational Motion Estimation. PAMI 2011.
30 HMM
This method uses optical flow information.
24.76 % 34.16 % 7.2 px 15.0 px 100.00 % 10 min 1 core @ 2.5 Ghz (C/C++)
Anonymous submission
31 DB-TV-L1
This method uses optical flow information.
code 30.75 % 39.13 % 7.8 px 14.6 px 100.00 % 16 s 1 core @ 2.5 Ghz (Matlab)
C. Zach, T. Pock and H. Bischof: A Duality Based Approach for Realtime TV- L1 Optical Flow. DAGM 2007.
32 GCSF
This method uses stereo information.
This method uses optical flow information.
33.23 % 41.74 % 7.0 px 15.3 px 48.27 % 2.4 s 1 core @ 2.5 Ghz (C/C++)
J. Cech, J. Sanchez-Riera and R. Horaud: Scene Flow Estimation by Growing Correspondence Seeds. CVPR 2011.
33 HAOF
This method uses optical flow information.
code 35.76 % 43.36 % 11.1 px 18.2 px 100.00 % 16.2 s 1 core @ 2.5 Ghz (C/C++)
T. Brox, A. Bruhn, N. Papenberg and J. Weickert: High accuracy optical flow estimation based on a theory for warping. ECCV 2004.
34 BERLOF
This method uses optical flow information.
37.59 % 45.20 % 8.5 px 16.2 px 15.26 % 0.231 s GPU @ 700 Mhz (C/C++) GeForce GTX 680
T. Senst, J. Geistert, I. Keller and T. Sikora: Robust Local Optical Flow Estimation using Bilinear Equations for Sparse Motion Estimation. 20th IEEE International Conference on Image Processing 2013.
35 RLOF
This method uses optical flow information.
code 38.51 % 46.04 % 8.7 px 16.5 px 14.76 % 0.488 s GPU @ 700 Mhz (C/C++) GeForce GTX 680
T. Senst, V. Eiselein and T. Sikora: Robust Local Optical Flow for Feature Tracking. TCSVT 2012.
36 PolyExpand
This method uses optical flow information.
47.54 % 53.95 % 17.2 px 25.2 px 100.00 % 1 s 1 core @ 2.5 Ghz (C/C++)
G. Farneback: Two-Frame Motion Estimation Based on Polynomial Expansion. SCIA 2003.
37 OCV-BM
This method uses optical flow information.
code 63.46 % 68.16 % 24.4 px 33.3 px 100.00 % 1.5 min 1 core @ 2.5 Ghz (C/C++)
G. Bradski: The OpenCV Library. Dr. Dobb's Journal of Software Tools 2000.
38 Pyramid-LK
This method uses optical flow information.
code 65.74 % 70.09 % 21.7 px 33.1 px 99.90 % 1.5 min 1 core @ 2.5 Ghz (Matlab)
J. Bouguet: Pyramidal implementation of the Lucas Kanade feature tracker. Intel 2000.
39 MEDIAN
This method uses optical flow information.
79.31 % 82.40 % 16.0 px 23.9 px 99.94 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
40 AVERAGE
This method uses optical flow information.
81.16 % 83.97 % 16.3 px 24.6 px 99.94 % 0.01 s 1 core @ 2.5 Ghz (C/C++)
This table as LaTeX

The settings column describes additional assumptions made / information used by the methods:

  • ms = motion stereo: Usage of the epipolar geometry to restrict the search problem to 1D

Related Datasets

  • HCI/Bosch Robust Vision Challenge: Optical flow and stereo vision challenge on high resolution imagery recorded at a high frame rate under diverse weather conditions (e.g., sunny, cloudy, rainy). The Robert Bosch AG provides a prize for the best performing method.
  • Image Sequence Analysis Test Site (EISATS): Synthetic image sequences with ground truth information provided by UoA and Daimler AG. Some of the images come with 3D range sensor information.
  • Middlebury Optical Flow Evaluation: The classic optical flow evaluation benchmark, featuring eight test images, with very accurate ground truth from a shape from UV light pattern system. 24 image pairs are provided in total.



eXTReMe Tracker